Natural Language in a voice service allows users to express their requests in their own words, without any dialogue constraints. Far beyond this freedom, which is an extra source of satisfaction for the user, Natural Language allows us to have a richer semantic base with which we can give more meaning to users’ requests.

Without Natural Language, we would have to write a traditional grammar that would need to anticipate all the phrases that the different users of a given service could freely pronounce to express a request. This would be too complex.

Telisma’s answer to this complexity is based on two complementary axes: Semantic Interpretation and Statistical Language Models.

The teliSpeech Recognizer Semantic Interpretation looks at the different elements of useful information (place, date, number…) contained in simple free speech. The application can simply ignore the form and way the information is expressed. This function paves the way for less directed and more sophisticated dialogues. Semantic Interpretation simplifies your application code, eliminates the necessary updates linked, on the one hand, to identical data coming from new speech, and on the other hand to the development of the service in a different language

The teliSpeech Statistical Language Model simplifies the use of grammars adapted to Natural Language; simultaneously improving time to market, and reducing the effort required to produce them. Bigrams, which are Statistical Language Models based on two-word sequences. They offer a solution for automatic qualification in call centres by allowing callers to speak freely when answering open questions in a given domain.

The implementation of teliSpeech bigrams combines classic recognition scores based on acoustics and the probability of a sequence of words in an expression. Its main benefit is to improve the global recognition performance, even for speakers who hesitate or who use idiomatic expressions. Telisma’s methodology for constructing teliSpeech bigrams requires a minimum amount of effort. It is supported by an implementation process and a set of tools which allow you to generate a bigram model: text corpus normalisation, creation of the bigram source and compilation, so that the resulting model can be used in the voice service.